An industry that already ran this experiment
On 5 September 2013, a working group convened under the FAA's Performance-based Aviation Rulemaking Committee and the Commercial Aviation Safety Team published Operational Use of Flight Path Management Systems, with 28 findings and 18 recommendations. The finding that mattered was not that automation had made flying less safe. It was that pilots now spent so little time flying manually that the skill required when automation disengaged had begun to decay. The recommendation followed directly: airline operating policy should encourage pilots to fly manually, at least periodically, across entire departure and arrival phases.
Note the shape of that answer. Aviation did not conclude that it had automated too much. It concluded that automating the routine had quietly defunded the practice through which competence was maintained, and that the practice therefore had to be reinstated deliberately, on a schedule, as policy rather than as preference. Nobody flies a departure by hand because the autopilot cannot. They fly it because the airline has decided that the capability must exist on the day it is needed.
Operations is now in the position aviation occupied in 2013, and the numbers are not ambiguous. At its Supply Chain Symposium in Orlando on 5 May 2026, Gartner reported that a survey of 509 supply-chain leaders conducted between July and October 2025 found 55% expecting a decline in entry-level hiring as a result of agentic AI. In the same session Gartner forecast that by 2030, 75% of supply-chain organisations that paused entry-level hiring in 2026 will pay premiums upward of 15% for early-career professionals. Simon Bailey, the analyst presenting, put the mechanism plainly: AI is not a plug-and-play replacement for people, and organisations that stop hiring and fail to develop early-career professionals will face pipeline gaps.
Two facts make this worse in the Southern Cone than the global averages suggest. ManpowerGroup's Talent Shortage work puts the share of Latin American employers struggling to fill roles at roughly 74%, and the scarcest logistics profile named in the regional trade press is precisely the one this article is about: the planner who combines analytical capability with operational feel. That combination is not taught. It is accumulated, case by case, in the layer of work now being automated first.
The demand for operational judgment is rising. The mechanism that produced it is being automated. Those are the same decision.
The Judgment Ladder
Socradata uses a three-rung model to describe how an operation manufactures the people who will later decide its hard cases. The rungs are sequential and cannot be skipped, which is the whole problem: automation enters at the bottom, where specification is easy and volume is high, and it removes the rungs in the order a person needs to climb them.
Seeing a large number of ordinary cases. This is how a person acquires the base rate: what a normal order profile looks like in March, how a reliable supplier behaves when it is late, which pick errors cluster and which are noise. The base rate is what makes an abnormal case visible at all, and it is not transferable by documentation, because it is a distribution rather than a rule. Rung 1 is also the cheapest work to specify and the first an agent absorbs. An operation can remove it in a single deployment and see nothing but improvement for two years.
Owning an outcome, being wrong, and carrying the correction. Judgment is largely a memory of costs, and a person acquires it by making a call that turns out badly in a way they cannot delegate. This rung does not disappear under agentic deployment; it thins, which is harder to notice. When the human's role becomes approving a recommendation, the felt cost of a wrong approval is lower, later and shared with a system, and the learning that a mistake used to produce is substantially weaker. Approving is not deciding, and an organisation that counts approvals as decisions will overestimate the depth of its bench.
Deciding well on the case that has no precedent, under time pressure, with incomplete information and a consequence attached. This is the only rung an agentic operation needs more of, because automating the routine raises the share of remaining human work that is genuinely hard. It is also the only rung that cannot be purchased. An adjudicator is a person who has climbed the first two rungs inside your operation, on your products, with your suppliers and your customers. The market can sell you a planner. It cannot sell you a planner who knows how your December behaves.
The spanning metric is the adjudicator replacement horizon: for each named exception class, the number of years until the operation can no longer staff it, given the current number of qualified people, the current attrition rate, and the current rate at which new people qualify. It is arithmetic, not sentiment, and it is deliberately uncomfortable, because in most operations that have automated the routine layer the qualification rate has quietly fallen to zero while attrition has not. A horizon of six years is a planning problem. A horizon of two is an operating one, and it will not announce itself until the person who holds the class resigns.
So what: automating the routine case is not only a cost decision. It is a decision to stop manufacturing the people who handle the cases you cannot automate, and nothing on the business case names that cost.
Three operating patterns
The patterns below are anonymised composites from operating work in Buenos Aires and the Southern Cone. Figures are illustrative targets rather than audited results. Each names the system, the decision loop, the rung being eroded and the practice that restores it.
Consumer-goods planning, CABA — the planner who never planned. System: the demand-planning and replenishment module of the ERP, with an agent proposing order quantities. Decision loop: weekly replenishment across an ambient portfolio. Rung eroded: Volume, then Consequence. Two years after deployment the team's acceptance rate on agent proposals sits above 95%, which reads as success until a promotion week arrives and nobody can say whether the proposal is wrong, because no planner in the room has ever built the distribution by hand. Practice: each planner plans one SKU family unassisted each month, before seeing the agent's answer, and the two are compared in a thirty-minute review where the disagreements are the agenda. The point is not to beat the agent. It is to retain the ability to know when it is wrong.
Third-party logistics operator, AMBA — the supervisor who starts at the top. System: WMS exception handling and labour allocation across a multi-client distribution centre. Decision loop: clearing pick shorts, address failures, cut-off conflicts and cold-chain holds. Rung eroded: Volume, sharply. The agent clears the standard exception classes, so a new supervisor's first unaided decision is an escalation, which is by construction rare, ambiguous and expensive. They are being asked to start on rung 3. Practice: a defined share of routine exceptions is routed to trainees under a time-boxed decision, with the agent's answer revealed the following day and the gap recorded per person rather than per shift. Cost of the reserve is budgeted as part of the automation, not charged to the training line where it will be cut in the first difficult quarter.
Mining-services group, Santiago de Chile — the buyer with no feel for the supplier. System: ERP purchase requisition and contract-price records against framework agreements. Decision loop: approval of maintenance, repair and operations requisitions. Rung eroded: Consequence. Routine requisitions clear automatically, so the category buyer touches only exceptions and sees supplier behaviour as a set of exception reports rather than as a pattern across hundreds of ordinary transactions. Three years later a framework agreement comes up for renegotiation and the buyer has price history but no accumulated read on which supplier bends and which walks. Practice: a rotating sample of routine requisitions is worked manually each quarter, and the buyer writes one paragraph per supplier per quarter on observed behaviour, which is the cheapest institutional memory an operation can build.
Budget the practice, not the intention
Start by computing the horizon, because it converts an abstract worry into a number an executive committee can act on. List the exception classes with financial, safety or regulatory consequence. For each, count the people currently qualified to decide it unaided, the annual attrition among them, and the number who qualified in the last twelve months. Where the third figure is zero and the second is not, the horizon is finite and short, and it can be stated in years. Most operations complete this in a fortnight, and the common result is a small number of classes carried by one or two people who are within a decade of retirement.
Then convert the finding into a practice reserve: a standing, budgeted, scheduled allocation of routine cases deliberately routed to humans, sized from the horizon rather than from a comfortable percentage. This is the aviation answer, and its logic is the same. The reserve is not a concession that the agent is unreliable. It is the recognition that capability which is never exercised is not capability, and that the day it is needed is not the day to discover this. Budget it inside the automation business case, where it will survive, rather than in the training line, where it will not.
So what: if your automation programme has no line item for deliberately doing work the agent could have done, you have not built an operating model. You have built a depreciation schedule for your own bench. From pilot to policy means the reserve is written into the operating procedure, not left to a supervisor's good intentions in a quarter when volume spikes.
Governance
One control, owned by the chief operating officer: an adjudicator roster with a practice reserve attached. For every exception class carrying consequence, the roster names the people qualified to decide it unaided, the date each last decided one without assistance, the reserve percentage routed to humans to keep the class current, and the computed horizon. It is reviewed quarterly alongside service level and cost per order, not annually with the talent plan, because it is an operating risk rather than a human-resources aspiration. The gate is simple and should be enforced at the deployment decision: no agent takes over a decision class until the reserve for that class is defined, sized and funded. An automation proposal that cannot say who will still be able to do the work in five years is not finished.
KPIs
Adjudicator replacement horizon: years until a named exception class cannot be staffed at current attrition and qualification rates. Baseline uncomputed; target above 5 years for every class with financial or safety consequence. Practice reserve rate: share of automatable volume deliberately routed to humans for skill maintenance. Baseline typically zero; target set per class from the horizon, not from a round number. Approval-to-decision ratio: share of human touches that are genuine unaided decisions rather than confirmations of an agent proposal. This is the rung-2 gauge, and a ratio near zero means the bench is thinning while the headcount is stable. Unassisted quality gap: difference between a trainee's unaided decision and the agent's on the same case, scored next day and tracked per person. Escalation resolution time by tenure band: rising resolution time among newer staff is the earliest visible signal that rung 1 has gone.
12-month roadmap
0–90: name the exception classes that carry consequence; compute the horizon for each and publish it internally, however poor; measure the approval-to-decision ratio on one automated loop and expect it to be worse than assumed. 90–180: define and fund the practice reserve for the two classes with the shortest horizon; put the reserve inside the automation business case; start the unassisted-then-compare routine on one team and record the gap per person. 180–360: make the reserve a deployment gate so no new decision class is automated without one; add the horizon and the approval-to-decision ratio to the operating scorecard beside service level; review the roster quarterly and re-qualify at least one new adjudicator per consequential class.
Capacity is easy to buy. Judgment has to be grown.
The prevailing account of agentic operations treats human judgment as a fixed input: the agent handles the routine, the person handles the exception, and the design question is where to draw the line. That account is incomplete in one important respect. Judgment is not a stock the organisation holds. It is a flow the organisation produces, and it is produced almost entirely by exposure to work that is now economically irrational to give a human.
The consequences of ignoring this are visible in adjacent industries. Klarna's publicly reported experience is the clearest: an automation programme announced in 2024 as handling work equivalent to hundreds of service roles, followed in 2025 by the chief executive's concession that the company had weighted efficiency and cost too heavily and that the resulting quality was not sustainable. The relevant lesson is not that automation failed. It is that the human-to-agent ratio has to be governed by an outcome, and that the recovery cost is asymmetric. Gartner's 2 September 2026 survey of 3,566 customers found only 27% willing to try a chatbot again after one negative experience. Trust leaks faster than it refills, and so does a bench.
Socradata transforms ERP, WMS and supply-chain data into predictive intelligence and governed operational decision systems — and in this domain, a governed system is one that has been designed to still work in five years, which means designing for the people who will have to run it then.
Compute your horizon before your next deployment
Every Wednesday, The Operational AI Dispatch takes one consequential AI signal and translates it into an operating model, a KPI set and a practical action plan for leaders running enterprise operations, ERP, WMS, supply chains and public systems. Published weekly by Socradata from Buenos Aires and New York.